Abstract:
With the advancement of Artificial Intelligence (AI) connected real-world, the accurate prediction of age and gender has become a concerning issue nowadays. In order to perform an independent analysis of bias in the UTKFace dataset, metadata of 23,645 pictures was obtained with the help of custom Python scripts and tabulated in Excel to be reviewed in detail. Analysis has revealed that where the number of male and female samples was almost equal, statistically significant, age difference between the groups of half a decade existed. The imbalance was also validated by a randomly chosen sample of 7,000 images, as the sample has some age-related bias that can be measured. EfficientNetB0, MobileNetV2, ShuffleNetV2, and the custom CNN are four deep-learning models that were trained to make sure there was a fair comparison. EfficientNetB0 gave the best results with over 90 percent accuracy in gender classification and credible age estimation. Further testing with confusion matrices, TPR comparisons and classification reports served to test potential bias in the algorithm. TPR difference between males and females was not large (0.90 and 0.91 respectively) which indicated that the general model structure and analysis adequately minimized the performance differences. Grad-CAM was used to deal with model interpretability, highlighting the most influential and important regions of the face in prediction. The significance of the method is that focusing with the forehead and mid - face region as per the human visual intuition which facilitated openness of judgment. Each stage of the working process was biased to check, statistically verified and interpretable.
